Multilingual Voices - Cartesia Docs

A multilingual voice speaks more than one language and sounds native in each, so one voice can serve customers in every language you support.

See what a voice can speak

In the voice library, hover over the flag in a voice’s language column to see the languages and accents it supports. In the API, call Get Voice and read the accents field:

// GET /voices/9626c31c-bec5-4cca-baa8-f8ba9e84c8bc
{
  "id": "9626c31c-bec5-4cca-baa8-f8ba9e84c8bc",
  "name": "Jacqueline",
  "accents": [
    { "accent": "general-american", "locale": "en-US", "is_native": true },
    { "accent": "hindi", "locale": "hi-IN", "is_native": false }
    // ...and every other accent this voice supports
  ]
  // ...plus description, gender, created_at, and the rest of the voice
}

In this example, Jacqueline can speak two accents: General American English and Hindi. Each accent belongs to a locale, and is_native: true marks the voice’s original accent.

Using a multilingual voice

Take Jacqueline from above and use that voice ID in a text-to-speech request. Most of the time, setting the locale you want spoken is all you need: the voice speaks that language in one of the accents it supports, and Cartesia normalizes the transcript for that locale.

locale needs sonic-3.6 or newer. On earlier models, use language with a base code like en instead. A request may set one or the other, never both.

Speak a language natively

In the example below, we set locale to en-US, and Jacqueline speaks American English like a local.

{
  "model_id": "sonic-3.6",
  "voice": "9626c31c-bec5-4cca-baa8-f8ba9e84c8bc",
  "transcript": "Thanks for holding. I've pulled up your account, and the last three payments went through without any trouble. The one from Thursday is still pending, so it should clear by tomorrow morning.",
  "locale": "en-US"
}

Add a deliberate accent

Setting accent is useful when you intentionally want the voice to sound accented, the way a non-native speaker would. Here the transcript is the same English as above, with accent set to hindi. Listen to the audio and notice how Jacqueline speaks English with a Hindi accent:

{
  "model_id": "sonic-3.6",
  "voice": "9626c31c-bec5-4cca-baa8-f8ba9e84c8bc",
  "transcript": "Thanks for holding. I've pulled up your account, and the last three payments went through without any trouble. The one from Thursday is still pending, so it should clear by tomorrow morning.",
  "locale": "en-US",
  "accent": "hindi"
}

Accent fallbacks

When you request a language this voice has no accent for, the voice will speak it with its original accent, General American English in the case of Jacqueline. Jacqueline doesn’t support a Japanese accent, so listen to the audio and notice how she speaks Japanese with an American accent:

{
  "model_id": "sonic-3.6",
  "voice": "9626c31c-bec5-4cca-baa8-f8ba9e84c8bc",
  "transcript": "お待たせいたしました。お客様のアカウントを確認しましたところ、直近三件のお支払いはすべて問題なく完了しております。",
  "locale": "ja-JP"
}

The same fallback applies at the locale level. Jacqueline’s only English accent is General American (en-US), so a request for en-GB falls back to her American English rather than failing. The audio below is identical to the en-US request above:

{
  "model_id": "sonic-3.6",
  "voice": "9626c31c-bec5-4cca-baa8-f8ba9e84c8bc",
  "transcript": "Thanks for holding. I've pulled up your account, and the last three payments went through without any trouble. The one from Thursday is still pending, so it should clear by tomorrow morning.",
  "locale": "en-GB"
}

To get British English, use a voice that lists en-GB in its accents field. Only pass an accent this voice supports. Jacqueline has no japanese accent in the accents field, so pinning one fails:

{
  "model_id": "sonic-3.6",
  "voice": "9626c31c-bec5-4cca-baa8-f8ba9e84c8bc",
  "transcript": "お待たせいたしました。お客様のアカウントを確認しましたところ、直近三件のお支払いはすべて問題なく完了しております。",
  "locale": "ja-JP",
  "accent": "japanese"
}

// HTTP 400
{
  "error_code": null,
  "message": "this voice does not support accent 'japanese'",
  "title": "Unsupported accent for voice",
  "request_id": "927f6ea3-22cf-426a-8767-446f41677e17"
}

Switching languages mid-conversation

Send one language per request. For example, send “For English, press one” and “हिंदी के लिए, दो दबाएँ” as two separate requests. Mixing languages inside a single generation works where it’s common, such as Hindi (Hinglish) and Tagalog (Taglish). Outside those cases the speech may sound accented.

Making your voice multilingual

You can turn your Instant Voice Clone multilingual beyond the original accent you cloned. Behind the scenes, Cartesia adapts the voice to every accent you add, keeping the speaker’s identity and the same voice ID. You can do this in the dashboard or the API:

Cartesia dashboard: find the voice in the voice library, open the More actions menu next to the voice, then select Manage voice > Supported languages > Add languages.
API: call Add Voice Accents with the accent ids you want this voice to speak.

A voice can support at most 10 accents. To free a slot, remove one with Delete Voice Accent.

Adding accents to a public voice or a Pro Voice Clone is not supported. Use the Localize Voice API instead, which creates a new voice with its own voice ID.

​See what a voice can speak

​Using a multilingual voice

​Speak a language natively

​Add a deliberate accent

​Accent fallbacks

​Switching languages mid-conversation

​Making your voice multilingual

See what a voice can speak

Using a multilingual voice

Speak a language natively

Add a deliberate accent

Accent fallbacks

Switching languages mid-conversation

Making your voice multilingual